First public near-lossless 5-bit Vision Transformer. Compressed from facebook/dinov2-large using the UltraCompress v3 codec -- the same pipeline that ships 22 transformer LLM packs at a mean PPL ratio of 1.005x.

This pack proves the codec generalizes beyond text transformers to vision architectures with zero format changes.

CLS-token cosine is the natural quality metric for DINOv2: it measures how closely the compressed model's learned representation matches the original across a diverse image set. 0.9988 means the compressed features are effectively interchangeable with the originals for downstream tasks (kNN classification, retrieval, linear probing).

For vision transformers, CLS-token cosine replaces perplexity (PPL) as the quality comparator. PPL is undefined for non-autoregressive models. The CLS-token cosine of 0.9988 on DINOv2-Large is comparable in quality preservation to the 1.004x PPL ratio achieved on similarly-sized LLMs (e.g., Qwen3-8B at 1.0044x).

Every layer file carries a per-tensor SHA-256 digest. To verify the SHA-256 digest of reconstruction:

UltraCompress v3 binary (.uc files). See the codec specification for format details. This pack uses the same v3 format as all LLM packs -- no format extensions were needed for vision transformers.

Apache 2.0 (same as base model).